Computer vision-based real-time monitoring system for inpatient behavior abnormalities

Abstract

The present application relates to the technical field of computer vision, and especially relates to a hospitalized patient behavior abnormality real-time monitoring system based on computer vision, which comprises a collection terminal, a preprocessing device, a modeling unit, a semantic analysis unit, a motion recognition engine, a research and judgment module and an interaction platform. By extracting skeleton key points to reconstruct postures, using semantic analysis to depict behavior evolution sequences, and introducing a shielding compensation mechanism to process complex scenes, the system can identify abnormalities such as falls and bed falls by calculating the deviation degree of real-time behavior from normal modes. The present application can realize all-weather automatic monitoring, significantly reduce the false alarm rate, and improve the safety management level of the ward.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a real-time monitoring system for abnormal behavior of hospitalized patients based on computer vision. Background Technology

[0002] Traditional ward rounds rely excessively on real-time human resources, making it difficult to achieve 24 / 7 coverage of patient conditions. This results in unforeseen risks such as falls, bed falls, and sudden physiological abnormalities not being addressed promptly. Existing technologies, when dealing with these issues, encounter problems with deep motion occlusion and difficulty in recognizing fine movements, leading to high false alarm and false negative rates when bed covering is in place or when medical staff frequently intervene.

[0003] Meanwhile, traditional methods focus on static determination of instantaneous actions and lack the ability to perform correlation analysis on long-term behavioral sequences of patients, making it difficult to effectively distinguish between normal activities and potential risks. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time monitoring system for abnormal behavior of hospitalized patients based on computer vision, so as to solve the technical problems mentioned in the background.

[0005] The present invention provides a real-time monitoring system for abnormal behavior of hospitalized patients based on computer vision, including a data acquisition terminal, a preprocessing device, a modeling unit, a semantic analysis unit, an action recognition engine, an analysis module, and an interactive platform;

[0006] The acquisition terminal is configured to capture video streams of hospitalized patients and their surrounding environment around the clock, and acquire raw image sequences containing patient position information, movement trajectory and medical device interaction status through a high-sensitivity optical sensor.

[0007] The preprocessing device is used to perform spatial domain denoising, brightness adaptive equalization, and temporal domain motion compensation on the received raw image sequence, and outputs a standardized candidate monitoring frame sequence.

[0008] The modeling unit is used to extract key feature points of the human body in the candidate monitoring frame sequence, establish a dynamic model, and realize real-time reconstruction of the patient's motion posture and generate dynamic trajectory data by performing optical flow analysis and geometric constraint matching on the feature points between adjacent frames.

[0009] The semantic analysis unit is configured to receive dynamic trajectory data and use a deep semantic extraction algorithm to encode the patient's behavior within a predetermined time span, thereby characterizing the temporal features of the patient's behavior evolution from stillness, turning over, sitting up to getting out of bed.

[0010] The action recognition engine is configured to process feature extraction of patients in occluded scenes. Through a spatial consistency reasoning mechanism, combined with the prior distribution pattern of the patient's historical posture, the position of the occluded joint is predicted and reconstructed.

[0011] The judgment module is based on a preset behavioral logic criterion library and identifies abnormal signs, including signs of falling and bed falls, by calculating the deviation between real-time behavior and normal mode.

[0012] The interactive platform is used to aggregate and analyze data and display the ward monitoring status in real time. When abnormal behavior is detected, alarm logic is triggered according to preset priority.

[0013] In some embodiments, the preprocessing apparatus includes at least a spatial denoising subunit, a brightness adaptive equalization subunit, and a background modeling subunit;

[0014] The spatial domain denoising subunit is configured to, for each pixel in the frame to be processed, find multiple reference blocks with similarity to the neighborhood of the pixel within a predetermined neighborhood window centered on the pixel, calculate the Euclidean geometric distance between each reference block and the target block, determine the contribution weight of the reference block in the weighted averaging process based on the distance, and determine the denoised pixel value by performing a nonlinear weighted summation on all reference blocks.

[0015] The background modeling subunit is configured to extract static background features from the image, construct and update the background field of the ward environment in real time through a Gaussian mixture model, and use a background subtraction algorithm to separate the patient's foreground target from the medical environment.

[0016] The spatiotemporal image feature preprocessing device is equipped with frame rate adaptive adjustment logic. When the patient is detected to be in deep sleep or stillness through inter-frame difference operation, the data sampling frequency is automatically reduced. When the patient is detected to be turning over or changing position, the sampling frequency is immediately increased to a predetermined high frequency level.

[0017] In some embodiments, the modeling unit pre-stores a human body topology knowledge base; the dynamic model includes at least twenty-four key dynamic nodes, including the head, torso, shoulders, elbows, wrists, hips, knees, and ankles.

[0018] The modeling unit also includes a center of gravity balance analysis unit, configured to evaluate the stability of the patient when standing or moving by calculating the projection position of the center of gravity of the reconstructed skeletal model on the three-dimensional projection plane in real time. If the distance between the center of gravity offset vector and the edge of the support surface is less than a preset safety threshold, it is determined that there is a risk of falling.

[0019] In some embodiments, the semantic analysis unit is a hierarchical feature fusion structure and has a behavior buffer with a configured sliding window mechanism.

[0020] In some embodiments, the semantic analysis unit is further configured to identify specific intentional behaviors of the patient, including the patient reaching for an object and attempting to find the call bell, and to trigger notification logic in advance based on the specific intentional behaviors.

[0021] In some embodiments, the motion recognition engine utilizes adversarial associative reasoning logic to infer the motion trend of the occluded part by analyzing the linkage characteristics of the unoccluded part and combining the kinematic prior distribution law of the patient's historical posture when a part of the patient's body is completely occluded.

[0022] The adversarial associative reasoning logic is implemented through the generator model and discriminator model maintained internally by the system.

[0023] The generator model attempts to complete and output the predicted pose coordinates of the occluded part based on the visible bone information of the unoccluded part. The discriminator model is responsible for evaluating whether the predicted pose conforms to the physiological and anatomical constraints of the human body. The optimal compensation result is output through the alternating evolution analysis between the generator and the discriminator.

[0024] In some embodiments, when calculating the deviation, the judgment module extracts the feature distribution function of the real-time behavior sequence in the feature space and calculates the Kolb-Leibler divergence between it and the probability density function of the pre-stored normal behavior mode; when the Kolb-Leibler divergence value exceeds a preset anomaly measurement threshold and the duration of the deviation on the time axis is greater than a preset debouncing period, an anomaly is determined to have occurred.

[0025] In some embodiments, the risk assessment module includes a self-learning optimization strategy, configured to dynamically adjust the sensitivity threshold for alarm triggering based on the age, severity of illness, and recovery stage of different patients, and to continuously correct the weight parameters in the judgment logic by analyzing historical false alarm records.

[0026] In some embodiments, the interactive platform includes a fault self-diagnosis subsystem that monitors the online status of each hardware terminal and the bandwidth usage of the data link in real time, and activates redundant links when hardware is found to be offline or resources are overloaded.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This invention enables 24 / 7 automatic monitoring of hospitalized patient behavior, replacing traditional manual inspections with computer vision technology, significantly reducing the workload of nursing staff, filling monitoring blind spots caused by limited medical resources, and improving the precision of ward safety management;

[0029] 2. This invention effectively solves the recognition problems caused by medical equipment occlusion and bedding coverage in complex ward environments by constructing dynamic posture skeleton modeling and tracking components and occlusion compensation engine, and significantly reduces the false alarm rate and missed alarm rate.

[0030] 3. The long-term behavioral sequence semantic analysis unit introduced in this invention breaks through the limitations of traditional single-frame image recognition, realizes the logical understanding of the evolution process of patient behavior, and can identify complex abnormal behaviors with time evolution characteristics, such as the warning signs of falling out of bed.

[0031] 4. This invention achieves accurate push and response of early warning information through the collaborative work of the judgment module and the interactive platform. Combined with self-learning optimization strategy and multi-level response mechanism, it can provide personalized safety protection solutions according to the needs of different patients, which greatly improves the intelligence level and response efficiency of clinical nursing.

[0032] 5. The present invention fully considers illumination compensation, privacy protection, multi-target recognition and multimodal data fusion in its system architecture, and has strong scene adaptability. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is an architecture diagram of the real-time monitoring system for abnormal behavior of hospitalized patients based on computer vision, according to the present invention.

[0035] Figure 2 This is a schematic diagram of the pretreatment device of the present invention. Detailed Implementation

[0036] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0037] Example 1

[0038] This embodiment provides a real-time monitoring system for abnormal behavior in hospitalized patients based on computer vision, including a data acquisition terminal, a preprocessing unit, a modeling unit, a semantic analysis unit, an action recognition engine, an analysis module, and an interactive platform. The data acquisition terminal is configured to capture video streams of hospitalized patients and their surrounding environment around the clock, acquiring raw image sequences containing patient position information, movement trajectories, and medical device interaction status through a high-sensitivity optical sensor. The preprocessing unit performs spatial domain denoising, adaptive brightness equalization, and temporal motion compensation on the received raw image sequences, and outputs a standardized candidate monitoring frame sequence. The modeling unit extracts key feature points of the human body from the candidate monitoring frame sequence, establishes a dynamic model, and achieves real-time reconstruction of the patient's movement posture and generates dynamic trajectory data by performing optical flow analysis and geometric constraint matching on feature points between adjacent frames.

[0039] The semantic analysis unit is configured to receive dynamic trajectory data and use deep semantic extraction algorithms to encode the patient's behavior within a predetermined time span, characterizing the temporal features of the patient's behavior evolution from stillness, turning over, sitting up to getting out of bed;

[0040] The action recognition engine is configured to process feature extraction for patients in occluded scenarios. Through a spatial consistency reasoning mechanism, combined with the prior distribution patterns of the patient's historical postures, it predicts and reconstructs the positions of occluded joints. The judgment module, based on a preset behavioral logic criterion library, identifies abnormal signs, including signs of falls and bed falls, by calculating the deviation between real-time behavior and normal modality. The interactive platform is used to summarize and analyze data and display the ward monitoring status in real time. When abnormal behavior is detected, it triggers alarm logic according to preset priorities.

[0041] Specifically, the acquisition terminal is deployed high on the ceiling or side wall of a designated area in the ward and is configured to capture video streams of hospitalized patients and their surrounding environment around the clock from multiple perspectives. The terminal integrates multiple highly sensitive complementary metal-oxide-semiconductor optical sensors, which can capture raw image sequences containing patient position information, movement trajectories, and medical device interaction status.

[0042] To address the complex lighting environment in hospital wards, the data acquisition terminal features infrared enhancement and adaptive light compensation logic. When the external natural light intensity is below a preset brightness threshold, the terminal automatically switches to infrared mode, emitting infrared light of a predetermined wavelength through an infrared supplemental lamp. An infrared filter, in conjunction with a photosensitive element, captures the patient's motion contours. The terminal supports multi-angle collaborative coverage. By using multiple acquisition points located on the ceiling and side walls of the ward, spatial geometric transformations are employed to map images from different perspectives onto a unified three-dimensional coordinate system, constructing a three-dimensional panoramic monitoring field and eliminating blind spots caused by medical equipment such as bedside tables and IV poles.

[0043] The preprocessing unit is connected to the acquisition terminal via Gigabit Ethernet or a dedicated high-speed data bus to receive and analyze the raw image sequence in real time. The unit includes a spatial denoising subunit, a brightness adaptive equalization subunit, a temporal motion compensation subunit, and a background modeling subunit. The spatial denoising subunit employs an adaptive nonlocal mean filtering algorithm to smooth Gaussian and salt-and-pepper noise in the raw image. The brightness adaptive equalization subunit calculates the histogram distribution of the image in real time and performs nonlinear contrast stretching on overly bright or dark areas to ensure that image details remain clear even under strong sunlight or dim lighting at night. The background modeling subunit is configured to extract static background features from the image, constructs and updates the background field of the ward environment in real time using a Gaussian mixture model, and uses a background subtraction algorithm to accurately separate the patient's foreground from the complex medical environment.

[0044] In addition, the device is equipped with frame rate adaptive adjustment logic. When the system detects that the patient is in deep sleep or stillness through inter-frame difference operation, it automatically reduces the data sampling frequency to reduce the backend computing load. Once the system detects that the patient has turned over or changed position, it immediately increases the sampling frequency to a predetermined high frequency level.

[0045] The modeling unit receives a preprocessed, standardized sequence of monitoring frames to extract key human feature points and build a multi-node dynamic model. This component has a pre-stored human topology knowledge base, covering 24 key dynamic nodes including the head, torso, shoulders, elbows, wrists, hips, knees, and ankles. By executing convolutional feature extraction logic, the component can identify the patient's skeletal position in each frame and achieve real-time reconstruction of the patient's motion posture by performing optical flow analysis and geometric constraint matching on feature points between adjacent frames.

[0046] To maintain monitoring continuity in multi-target intervention scenarios (such as ward rounds by medical staff), the dynamic posture skeletal modeling and tracking component introduces a locking mechanism based on deep feature re-identification. This mechanism extracts the patient's posture feature vectors and, through calculation of the Euclidean distance between feature vectors during overlapping, occlusion, or entry and exit of multiple individuals, ensures that the system always locks onto the target patient, preventing the monitoring trajectory from drifting between different individuals. Simultaneously, the component includes a center of gravity balance analysis unit, which assesses the stability of the patient when standing or moving by calculating the projected position of the center of gravity of the reconstructed skeletal model on the 3D projection plane in real time. If the distance of the center of gravity offset vector relative to the edge of the support surface is less than a preset safety threshold, it is determined that there is a risk of fall.

[0047] The semantic analysis unit is configured to receive dynamic trajectory data and deeply encode patient behavior using a hierarchical feature fusion structure. This unit not only focuses on instantaneous pixel-level changes but also emphasizes analyzing the logical patterns of high-dimensional skeletal topological features evolving over time. It can transform disparate movements, such as raising an arm or kicking a leg, into clinically semantic behavioral sequences, such as attempts to grasp the bed rails or tremors caused by a violent cough. Regarding the risk of falls from bed, this unit analyzes the rate of change in the spatial distance between the patient's extremities and the bed edge, combined with the acceleration vector of the center of gravity shift, to assess the patient's behavioral evolution from rest to sitting up and then getting out of bed.

[0048] Furthermore, the unit has a specific intent recognition function, which can recognize specific action sequences such as a patient reaching for an object or searching for a call bell, thus realizing the transformation from passive response to active perception.

[0049] The motion recognition engine is specifically optimized for extreme scenarios unique to hospital wards, such as blanket covering and medical staff shielding. The engine uses adversarial associative reasoning logic. When part of the patient's body is completely covered by blankets, the system uses the motion linkage characteristics of the uncovered parts and the kinematic prior distribution of the patient's historical postures to perform spatial consistency reasoning prediction of the position of the occluded joints.

[0050] Meanwhile, the engine has a specific medical operation filtering function. By recognizing the color characteristics of the standard clothing of medical staff and the specific action patterns of multiple people working together, it automatically marks normal nursing operations, dressing changes or bedside examinations as whitelisted events, effectively filtering false alarms caused by third-party intervention.

[0051] The analysis module, based on a pre-defined behavioral logic criterion library, performs multi-dimensional comparative analysis of identified behavioral sequences with pre-stored normal physiological activity patterns. This module supports collaborative judgment logic, integrating real-time physiological data from existing heart rate monitoring, respiratory monitoring, or mattress pressure sensors within the ward through standardized interfaces. When visually identified abnormal movements highly overlap with abnormal heart rate waveforms detected by physiological sensors on the time axis, the system automatically increases the confidence level of the warning.

[0052] This module also includes a self-learning optimization strategy, which can dynamically adjust the sensitivity threshold of alarm triggering based on the age and severity of the illness of different patients, and continuously correct the weight parameters in the judgment logic by analyzing historical false alarm records. In addition, this module allows nursing experts to customize abnormal rules and define specific risk action combinations through a graphical interface.

[0053] The interactive platform aggregates the analysis results from each module and provides a human-computer interaction interface. When abnormal behavior is detected, the platform sends an immediate alarm to the corresponding nursing terminal according to a preset priority. The alarm information includes not only a text description but also automatically extracts 10-second video clips before and after the abnormality for medical staff to review.

[0054] The platform also supports a multi-level response mechanism. When high-risk warning signs are detected, it can control the ward lighting system to increase brightness and activate the bedside voice interaction module to play reminder messages.

[0055] Example 2

[0056] This embodiment mainly describes the algorithm details of the system's data processing logic. In the preprocessing device, the specific logic for spatial domain denoising is configured as follows: for each pixel in the frame to be processed, within a predetermined neighborhood window centered on the pixel, multiple reference blocks with high similarity to the neighborhood of the pixel are searched; the Euclidean geometric distance between each reference block and the target block is calculated, and the contribution weight of the reference block in the weighted averaging process is determined based on the distance; finally, the denoised pixel value is determined by performing a nonlinear weighted summation on all reference blocks.

[0057] In the modeling unit, the logic for extracting skeletal keypoints is configured as follows: Standardized monitoring frames are input into a pre-defined deep neural network model, which extracts high-dimensional feature maps through multi-layer convolution and pooling operations. Multiple confidence maps are generated on the feature maps, each representing the probability distribution of a specific joint point appearing at that spatial location. Simultaneously, a partial correlation field is generated to describe the spatial connection directions between different joint points. By performing local maxima search on the confidence maps and combining this with greedy parsing of the correlation field, discrete joint points are connected into a complete skeletal link.

[0058] In the semantic analysis unit, the system constructs a behavior buffer with a sliding window mechanism, the length of which corresponds to a predetermined duration. At each sampling period, the extracted skeletal feature vector, including joint coordinates, limb segment angles, and angular velocities, is pushed into the buffer. A recurrent neural network structure is used to perform hidden layer state mapping on the temporal data within the buffer, generating semantic vectors that characterize the dynamic evolution of the behavior. These semantic vectors are then fed into a classification layer, where the overlap between their projections and the predefined behavior category centers in the semantic space is calculated to determine the label of the current behavior and its probability of occurrence.

[0059] In the action recognition engine, the implementation logic of adversarial association reasoning is as follows: the system internally maintains a generator model and a discriminator model. The generator model attempts to complete and output the predicted pose coordinates of the occluded part based on the visible skeletal information of the unoccluded part; the discriminator model is responsible for evaluating whether the predicted pose conforms to the physiological and anatomical constraints of human movement (such as the elbow joint not bending in the opposite direction). Through alternating evolutionary analysis between the generator and the discriminator, an optimal compensation result that conforms to both the current visual observation and the laws of biodynamics is finally output.

[0060] In the analysis module, the deviation calculation process is configured as follows: extract the feature distribution function of the real-time behavior sequence in the feature space, and calculate the Koolbek-Leibler divergence between it and the probability density function of the pre-stored normal behavior mode. When the divergence value exceeds a preset anomaly measurement threshold, and the duration of the deviation on the time axis is greater than a preset debouncing period, the system determines that an anomaly has occurred, and determines the specific type of anomaly based on the component weights of the deviation in different dimensions.

[0061] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0062] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A real-time monitoring system for abnormal behavior in hospitalized patients based on computer vision, characterized in that, It includes a data acquisition terminal, a preprocessing device, a modeling unit, a semantic analysis unit, an action recognition engine, an analysis module, and an interactive platform; The acquisition terminal is configured to capture video streams of hospitalized patients and their surrounding environment around the clock, and acquire raw image sequences containing patient position information, movement trajectory and medical device interaction status through a high-sensitivity optical sensor. The preprocessing device is used to perform spatial domain denoising, brightness adaptive equalization, and temporal domain motion compensation on the received raw image sequence, and outputs a standardized candidate monitoring frame sequence. The modeling unit is used to extract key feature points of the human body in the candidate monitoring frame sequence, establish a dynamic model, and realize real-time reconstruction of the patient's motion posture and generate dynamic trajectory data by performing optical flow analysis and geometric constraint matching on the feature points between adjacent frames. The semantic analysis unit is configured to receive dynamic trajectory data and use a deep semantic extraction algorithm to encode the patient's behavior within a predetermined time span, thereby characterizing the temporal features of the patient's behavior evolution from stillness, turning over, sitting up to getting out of bed. The action recognition engine is configured to process feature extraction of patients in occluded scenes. Through a spatial consistency reasoning mechanism, combined with the prior distribution pattern of the patient's historical posture, the position of the occluded joint is predicted and reconstructed. The judgment module is based on a preset behavioral logic criterion library and identifies abnormal signs, including signs of falling and bed falls, by calculating the deviation between real-time behavior and normal mode. The interactive platform is used to aggregate and analyze data and display the ward monitoring status in real time. When abnormal behavior is detected, alarm logic is triggered according to preset priority.

2. The system according to claim 1, characterized in that, The preprocessing device includes at least a spatial denoising subunit, a brightness adaptive equalization subunit, and a background modeling subunit; The spatial domain denoising subunit is configured to, for each pixel in the frame to be processed, find multiple reference blocks with similarity to the neighborhood of the pixel within a predetermined neighborhood window centered on the pixel, calculate the Euclidean geometric distance between each reference block and the target block, determine the contribution weight of the reference block in the weighted averaging process based on the distance, and determine the denoised pixel value by performing a nonlinear weighted summation on all reference blocks. The background modeling subunit is configured to extract static background features from the image, construct and update the background field of the ward environment in real time through a Gaussian mixture model, and use a background subtraction algorithm to separate the patient's foreground target from the medical environment. The spatiotemporal image feature preprocessing device is equipped with frame rate adaptive adjustment logic. When the patient is detected to be in deep sleep or stillness through inter-frame difference operation, the data sampling frequency is automatically reduced. When the patient is detected to be turning over or changing position, the sampling frequency is immediately increased to a predetermined high frequency level.

3. The system according to claim 1, characterized in that, The modeling unit contains a pre-stored human topology knowledge base; the dynamic model includes at least twenty-four key dynamic nodes, including the head, torso, shoulders, elbows, wrists, hips, knees, and ankles. The modeling unit also includes a center of gravity balance analysis unit, configured to evaluate the stability of the patient when standing or moving by calculating the projection position of the center of gravity of the reconstructed skeletal model on the three-dimensional projection plane in real time. If the distance between the center of gravity offset vector and the edge of the support surface is less than a preset safety threshold, it is determined that there is a risk of falling.

4. The system according to claim 1, characterized in that, The semantic analysis unit is a hierarchical feature fusion structure and has a behavior buffer with a configured sliding window mechanism.

5. The system according to claim 4, characterized in that, The semantic analysis unit is also configured to recognize specific intentional behaviors of the patient, including the patient reaching for an object and attempting to find the call bell, and to trigger notification logic in advance based on the specific intentional behaviors.

6. The system according to claim 1, characterized in that, The motion recognition engine uses adversarial associative reasoning logic to infer the motion trend of the occluded part when a part of the patient's body is completely occluded. This is achieved by analyzing the linkage characteristics of the unoccluded part and combining the kinematic prior distribution law of the patient's historical posture. The adversarial associative reasoning logic is implemented through the generator model and discriminator model maintained internally by the system. The generator model attempts to complete and output the predicted pose coordinates of the occluded part based on the visible bone information of the unoccluded part. The discriminator model is responsible for evaluating whether the predicted pose conforms to the physiological and anatomical constraints of the human body. The optimal compensation result is output through the alternating evolution analysis between the generator and the discriminator.

7. The system according to claim 1, characterized in that, When calculating the deviation, the judgment module extracts the feature distribution function of the real-time behavior sequence in the feature space and calculates the Kolb-Leibler divergence between it and the probability density function of the pre-stored normal behavior mode. When the Kolb-Leibler divergence value exceeds the preset anomaly measurement threshold and the duration of the deviation on the time axis is greater than the preset de-jittering period, an anomaly is determined to have occurred.

8. The system according to claim 7, characterized in that, The risk assessment module includes a self-learning optimization strategy, configured to dynamically adjust the sensitivity threshold for alarm triggering based on the age, severity of illness, and recovery stage of different patients, and continuously correct the weight parameters in the judgment logic by analyzing historical false alarm records.

9. The system according to claim 1, characterized in that, The interactive platform includes a fault self-diagnosis subsystem, which monitors the online status of each hardware terminal and the bandwidth usage of the data link in real time, and activates redundant links when hardware is found to be offline or resources are overloaded.

Images